Computing operation procedures for chemical plants using whole-plant simulation models

摘要

Chemical plants are complex dynamical systems. Optimising plant operation for non-stationary scenarios, such as changing the output product and recovering from abrupt disturbances, is challenging because a chemical plant has many operation points and complex responses. A plant simulator can be used to compute the optimal procedures. However, because of modelling errors or contingent changes in the external conditions, such as weather and feed purity, there exist gaps between the behaviour of a simulator and that of a real plant. This poses another challenge in a simulator-based approach, which adds to the computational complexity of the problem. In this study, we propose a simulator-based approach for optimising chemical plant operations using deep reinforcement learning and knowledge-based automated reasoning. Specifically, a reinforcement learning agent is trained on a whole-plant simulator with a policy gradient algorithm, using automated reasoning to narrow down the action space of the agent. To maintain the optimality of the procedures in a real plant, a simple method for the state and parameter estimation of the system at run time is introduced. This method can improve the accuracy of the response prediction model (i.e. the plant simulator) on which the agent depends. The presented method is evaluated on a real chemical distillation plant. The experimental results indicate that the proposed approach consumed only half the time and steam (heat energy) in comparison with that in the case of human-emulated procedures.